A comparative study on crash-influencing factors by facility ......A comparative study on...
Transcript of A comparative study on crash-influencing factors by facility ......A comparative study on...
A comparative study on crash-influencing factors by facility typeson urban expressway
Yong Wu • Hideki Nakamura • Miho Asano
Received: 1 July 2013 / Revised: 15 October 2013 / Accepted: 18 October 2013 / Published online: 30 November 2013
� The Author(s) 2013. This article is published with open access at Springerlink.com
Abstract This study aims at identifying crash-influencing
factors by facility type of Nagoya Urban Expressway,
considering the interaction of geometry, traffic flow, and
ambient conditions. Crash rate (CR) model is firstly
developed separately at four facility types: basic, merge,
and diverge segments and sharp curve. Traffic flows are
thereby categorized, and based on the traffic categories, the
significances of factors affecting crashes are analyzed by
principal component analysis. The results reveal that, the
CR at merge segment is significantly higher than those at
basic and diverge segments in uncongested flow, while the
value is not significantly different at the three facility types
in congested flow. In both un- and congested flows, sharp
curve has the worst safety performance in view of its
highest CR. Regarding influencing factors, geometric
design and traffic flow are most significant in un- and
congested flows, respectively. As mainline flow increases,
the effect of merging ratio affecting crash is on the rise at
basic and merge segments as opposed to the decreasing
significance of diverging ratio at diverge segment. Mean-
while, longer acceleration and deceleration lanes are
adverse to safety in uncongested flow, while shorter
acceleration and deceleration lanes are adverse in con-
gested flow. Due to its special geometric design, crashes at
sharp curve are highly associated with the large centrifugal
force and heavy restricted visibility.
Keywords Crash-influencing factors � Crash rates �Principal component analysis � Facility types � Urban
expressway
1 Introduction
Improving traffic safety is a worldwide issue to be relieved
urgently. Crash characteristics and their influencing fac-
tors, as the theoretical basis for safety improvement, may
provide direction for policies and countermeasures aimed
at smoothing hazardous conditions. For a better under-
standing of crash-influencing factors, researchers have
continually sought ways through an extensive array of
approaches, and the most prominent one is crash data
analysis [1]. The conventional approaches have established
statistical links between crash rate (CR) and its explanatory
factors [2, 3]. In the analyses, traffic flows are generally
represented by low-resolution data that is collected at a
highly aggregated level, e.g., hourly or daily flows. Geo-
metric features are primarily considered the hierarchy of
radius or slope [4, 5]. Meanwhile, several studies have
suggested that crashes are associated with the interaction of
geometry, traffic flow, and ambient conditions [6]. How-
ever, most existing studies investigated the factors indi-
vidually and the related CR models were developed based
on single factor only. As a result, it is inadequate to identify
the nature of individuals through aggregated analysis only,
since the conditions preceding individual crashes are vir-
tually different from each other [1].
Considering the insufficiency of CR analysis above,
some studies have tried to identify crash characteristics at
individual level, in an effort to predict crash risk on a real
time basis [7–9]. Through these studies, the effect of traffic
flow on crash risk has been well analyzed. In theory, the
concept of real-time crash prediction exhibits huge promise
for the application of proactive traffic management strate-
gies for safety.
However, the combined effects of geometry, traffic flow,
and ambient conditions on crashes still have not been well
Y. Wu (&) � H. Nakamura � M. Asano
Department of Civil Engineering, Nagoya University,
C1-2(651) Furo-cho, Chikusa-ku, Nagoya 464-8603, Japan
e-mail: [email protected]
123
J. Mod. Transport. (2013) 21(4):224–235
DOI 10.1007/s40534-013-0024-9
anayzed through the above studies. Furthermore, these
papers primarily developed crash model for the whole
traffic conditions, which may conflict with the fact that the
influence of traffic flow on crashes may vary when traffic
conditions change. In addition, even if crash characteristics
are found out to be dependent on facility type that is
composed of uniform segment individually, e.g., basic,
merge, and diverge segments [2], the existing studies are
focused on the entire route of intercity expressway without
segmentation.
Another cause for the limited predictive performance of
existing models is the inadequacy of analytic process [10].
As for statistical methods, the significance and indepen-
dence of explanatory variables should be identified in
advance for the reliability of statistics. Whereas, many
previous studies paid little attention to this point and
incorporated the potential influencing factors into crash
modeling directly.
Urban expressway is one common type of separated
highway with full control of access in large cities in Japan.
Generally, it is composed of various facility types where
geometric features and traffic characteristics are often
different from each other. Correspondingly, crash charac-
teristics and their influencing factors may also be different
by facility type. In the meantime, compared to intercity
expressway, crash characteristics and their related influ-
encing factors of urban expressway are different [11].
Necessarily, urban expressway deserves to be analyzed
independently and its crash characteristics should be
identified based on specific facility type.
Given the problems of existing studies, the objective of
this paper is to investigate crash characteristics based on
CR models and their influencing factors by facility type on
urban expressway. Meanwhile, the causes are identified by
considering the interaction of geometry, traffic flow, and
ambient conditions. Besides, geometric features are iden-
tified considering the driver-vehicle-roadway interaction.
The significances of these factors affecting crashes are
compared at different facility types using principal com-
ponent analysis (PCA). Their influencing mechanisms are
further discussed. In essence, this study can be regarded as
a proactive analysis for crash risk prediction model in the
future.
2 Study sites and datasets
2.1 Study sites
The test bed of this study is Nagoya Urban Expressway
network (NEX) as shown in Fig. 1. Up to December 31,
2009, this network was about 69.2 km 9 2 (two direc-
tions) in total length with over 250 ultrasonic detectors
installed with an average spacing of 500 m (varied in
250–750 m) on mainline. Most routes are 4-lane road-
ways (2-lane/dir), except the inner ring (Route no. R) that
is one-way roadway and where the number of lanes dif-
fers (2–5) with the change of ramp junctions. In the
limited areas, such as the links of other routes to the inner
ring, small curves are designed. In this network, two
recurrent bottlenecks are located along Odaka line (Route
no. 3).
Five databases are used in this study; (1) crash records
with the occurrence time in minutes, the location in 0.1 km
and the weather and pavement conditions; (2) detector data
including traffic volume q, average speed v, and occupancy
occ per 5 min; (3) geometric design and the location of
detector in 0.01 km; (4) traffic regulation records for
incidents (e.g., crash, working, and inclement weather)
including the locations and periods of temporal lane and
cross-section closures; and, (5) daily sunrise and sunset
time records in Nagoya. Here, it is worth noting that
detector data are processed for the whole cross section of
each direction. The period of the data above is for 3 years
(2007–2009) except for those on Kiyosu line (Route no. 6)
that was opened from December 1, 2007.
Fig. 1 Schematic map of NEX network (2009) (Source Nagoya
Expressway Public Corp., modified by authors)
Comparative study on crash-influencing factors 225
123J. Mod. Transport. (2013) 21(4):224–235
2.2 Segmentation of facility types
Basic segment is extracted outside the 500 m up- and
down-stream of ramp junctions considering the experience
in Japan [12]. Correspondingly, merge or diverge segments
are regarded as the sections inside the 500 m up- and
down-stream of on- and off-ramps, respectively. The seg-
mentation methods are shown in Fig. 2. Other than these
segments, there is a special geometric design in NEX,
curve with small radius. Figure 3 explains CR statistics
dependent on radius. Obviously, compared to other seg-
ments, much higher CR exists in the curves with radius
smaller than 100 m. Thus, these curves are defined as sharp
curves and regarded as another distinct facility type of
NEX. Given the limitation of segment samples available,
basic, merge, and diverge segments, and sharp curve will
be analyzed in this study.
The cross sections of inner ring are diverse and the
length of individual layouts, i.e., 2-, 3-, 4-, or 5-lane, is not
enough to be separately analyzed. Meanwhile, all of the
sharp curves along Inner ring are 2-lane roadway. In this
regard, only 2-lane segments are analyzed in this study and
the geometric statistics by facility types are summarized in
Table 1.
3 Methodology
3.1 Data extraction
3.1.1 Detector data
In principle, detectors can count the number of vehicles at
their locations only. In such case, the ‘‘coverage area’’ of
detector is defined for estimating traffic conditions at crash
locations through detector data. At basic segment, the
boundary of two consecutive coverage areas is defined at
the midpoint between two neighboring detectors. At merge
and diverge segments, it is bounded at the ramp-junction
point, and one segment can be divided into up- and down-
stream areas. Each sharp curve can be matched with a
single detector. Note that the time of crash is recorded by
road administrators after the crash occurrence. In reality, it
does not correspond to occurrence time exactly. For this
reason, data within small time before crashes should be
rejected to avoid mixing up crash-influencing and crash-
influenced data. Therefore, the latest data at least 5 min
before the recorded time are accepted after the exclusion of
invalid data and the data within lane and section-closure
intervals in advance.
3.1.2 Geometric features
Design consistency is the conformance of geometry of a
highway with driver expectancy, and its importance and
significant contribution to road safety is justified by
understanding the driver–vehicle–roadway interaction [13]
that may vary at individual locations in nature. In this
regard, geometric variation in the upstream of crash loca-
tion is proposed to reflect the effect of geometry on cra-
shes. Considering the length of detector coverage area, the
following variables in 500 m distance are extracted [12].
Basic segment
Basic segment
Merge segment
Diverge segment
Tight curve
0.5 km 0.5 km
0.5 km
0.5 km
0.5 km
0.5 km
R<100 m
Fig. 2 Segmentation of facility types
0
5
10
15
20
25
10 100 1000 10000 100000
Cra
sh ra
te (
cras
h/10
6 VK
MT
)
Radius (m)
Much higher CR
Fig. 3 Distribution of crash rate to radius
Table 1 Geometric statistics of facility types
Facility type No. of segments Total length (km)
Basic segment 38 56.6
Merge segment 28 20.9
Diverge segment 35 25.2
Sharp curve 9 2.5
226 Y. Wu et al.
123 J. Mod. Transport. (2013) 21(4):224–235
(1) Variation in road elevation h between the crash
location and its 500 m upstream, and the maximum ele-
vation difference H in this 500 m distance (Fig. 4).
(2) Horizontal displacement S. Radius is impossible to
describe a section composed of various curves. On the
other hand, centrifugal force is also associated with the
horizontal displacement s in the direction of tangent to the
curve j (Fig. 5). In such case, S in the 500 m distance (Rsj)
is adopted and calculated by the following equations.
hj ¼Lj
Rj
ð0 [ ; hj � p=2Þ; ð1Þ
sj ¼ Rjð1 � cos hjÞ; ð2Þ
where j is the ID of curve. Rj, hj, Lj, and sj correspond to the
radius, central angle, arc length, and horizontal displace-
ment of curve j, respectively.
(3) Index of centrifugal force ICF. Speed v always has a
square relation with centrifugal force. This study designs
ICF (ICF = Sv2) to reflect the combined effect of speed
v and horizontal displacement S, while it is not centrifugal
force.
(4) Index of space displacement ISD. ISD (ISD = SH) is
used to reveal the comprehensive geometric features
induced by horizontal and vertical variation in this study.
The geometric data above are collected every 0.1 km as
crash is recorded in a unit of 0.1 km. Besides, these data
are also extracted at the location of detector that is the
common link between crash and detector data. Table 2
summarizes the process of data collection.
3.1.3 Ambient conditions
Common, prevailing, and uncontrolled environment and
weather conditions are defined as ambient conditions. They
are (1) ambient light classified into daytime and nighttime,
which are the period from sunrise to sunset and from sunset
to sunrise, respectively; (2) sunny, cloudy, and rainy
weather conditions at the time of crash; (3) dry and wet
pavement conditions at the location of crash; and, (4) day
type on crash days including holiday and weekday. Here,
holiday includes all weekends, and all national and tradi-
tional holidays like the Golden Week in May and the Obon
Week in August in Japan.
3.1.4 Data matching
The related detector data, geometric features, and ambient
conditions for individual crashes are matched as exempli-
fied in Table 3. The crashes matching with invalid detector
data and within lane and cross-section closure intervals are
excluded in advance. As a result, a total of 1,591 crashes
remain for the following analysis.
h(>0) H
h(>0) H
500m upstream
Crash location
Crash location
500m upstream
Fig. 4 Variation in road elevation
sj
S
Lj
Lj+1
Sj+1
Rj
Rj
Rj+
1
θj
θ j+1
Fig. 5 Horizontal displacement
Table 2 Example of geometric variations collection
Route
no.
Direction* Kilopost** h (m) H (m) S (m) ISD
(m2)
1 SB 0.0 -4.63 5.49 0.78 4.30
1 SB 0.1 -7.90 8.49 3.91 33.2
1 SB 0.2 -10.6 11.5 6.08 69.9
1 SB 0.21 -11.5 11.8 8.88 104.7
1 SB 0.3 -15.3 15.3 9.60 146.9
1 SB – – – – –
1 SB 6.4 10.2 10.9 5.15 56.1
*SB South-bound
** 0.21: the Kilopost of detector #0101
Comparative study on crash-influencing factors 227
123J. Mod. Transport. (2013) 21(4):224–235
3.2 Classification of traffic conditions
Congested flow, characterized by traffic oscillation, has
different features from uncongested flow. It is necessary to
make a distinction between two traffic regimes. Figure 6
shows the traffic volume–speed diagram at Horita on-ramp
junction, one typical bottleneck in NEX. The speed of
60 km/h, corresponding to maximal flow is defined as the
critical speed vc that is used for classifying un- and con-
gested flows [2, 14]. Besides, the corresponding value at
another bottleneck (Takatsuji on-ramp junction) is also
found out around 60 km/h.
The value of 60 km/h would be regarded as the related
index at basic and diverge segments, since no bottleneck can
be virtually found at both segments in NEX. At sharp curve, a
threshold speed of 45 km/h is selected in general for clas-
sifying two traffic regimes based on traffic flow-speed dia-
gram at Tsurumai curve (Fig. 7). The value is further
checked at other sharp curves, and it is found out to be reli-
able for classifying un- and congested flows basically.
To reflect the variation in traffic characteristics, each
traffic regime is further sub-classified. It is evident that
speed has a high variance at low flow rates (see Figs. 6 and
7). Besides, occupancy is not a commonly used index.
Thus, traffic density k calculated by Eq. (3) is proposed to
be the measure of effectiveness to further classify the
traffic conditions. In view of the number of crash samples
available, the aggregation intervals of k are set as 10 and
30 veh/km for un- and congested flows, respectively.
kei ¼12 � qi
vi
; ð3Þ
where qi and vi denote traffic flow and average speed in
5 min # i, respectively. kei corresponds to the calculated
traffic density in this 5 min.
3.3 Calculation of crash rate (CR)
CR for traffic condition n can be calculated by the fol-
lowing equation:
CRn ¼ NOCn � 106
PQnlLl
; ð4Þ
where n and l are the ID of traffic condition and coverage
area, respectively; NOCn is the number of crashes for
traffic condition n. QnlLl is the value of vehicle kilometers
traveled (VKMT) in detector coverage area l for traffic
condition n.
Table 3 Examples of data matching for individual crashes
Crash
ID
Traffic characteristics Geometric features Ambient conditions
q (veh/
5 min)
v (km/h) MR DR Facility
typeah (m) H (m) ICF
(km3/h2)
ISD
(m2)
LAb
(m)
LDb
(m)
Light Weather Pavement Day type
1 139 88.7 – – B 1.5 1.5 0 0 – – Day Sunny Dry Holiday
2 96 95.0 – 0.02 D 0.5 1.3 302 45 – 220 Day Sunny Dry Weekday
3 29 80.3 – – S 10.3 12.7 126 249 – – Night Cloudy Wet Holiday
4 154 82.9 0.07 – M 1.5 1.5 0 0 200 – Day Cloudy Dry Weekday
B, M, D, and S basic, merge, and diverge segments, and sharp curve, respectively, LA length of acceleration lane at merge segment, LD length of
deceleration lane at diverge segment
0 10 20 30 40 50 60 70 80 90
100 110 120
0 25 50 75 100 125 150 175 200 225 250 275 300 325 350
Ave
rage
spe
ed (k
m/h
)
Traffic volume (veh/5-min)
Critical speed: 60km/h
High speed variability at low flow rates
Max. flow
Uncongested flow
Congested flow
Fig. 6 q–v diagram at Horita on-ramp junction
0
10
20
30
40
50
60
70
80
90
0 25 50 75 100 125 150 175 200 225 250 275 300
Ave
rage
spe
ed (k
m/h
)
Traffic volume (veh/5-min)
Threshold speed: 45km/h
Uncongested flow
Congested flow
High speed variability at low flow rates
Fig. 7 q–v diagram in Tsurumai curve
228 Y. Wu et al.
123 J. Mod. Transport. (2013) 21(4):224–235
3.4 Principal component analysis (PCA)
PCA is a powerful tool for reducing a large number of
observed variables into a small number of artificial vari-
ables that account for most of the variance in the dataset
[15]. In general, through orthogonal transformation, a set of
observations of possibly correlated variables can be con-
verted into a set of values of linearly uncorrelated vari-
ables. Those converted values are defined as principal
components. Technically, a principal component can be
regarded to be a linear combination of optimally weighted
observed variables [15]. As a result, the components are
ranked in the order of accounting amount of total variance
in the observed variables. Then, two criteria are generally
available to select the number of component extracted: (1)
80 % rule, the extracted components should be capable to
explain at least 80 % of the variance in the original dataset.
(2) Eigen value rule, only components whose eigen values
are over 1.0 can be retained.
4 Crash rate estimation models
In the following, the differences of crash characteristics by
facility type are investigated by comparing CR models
based on traffic conditions.
4.1 Uncongested flow
Figure 8 gives the CR tendency following traffic density
k by facility type in uncongested flow. It is evident that
sharp curve has a special characteristic compared to other
segments. Its CR is the highest among four facility types at
low-density stages. Then, the value follows a decreasing
tendency to k. In contrast, the CR at other segments
increases as k increases. Such phenomenon may be related
to the design of small radius for sharp curve. Such geo-
metric design can result in high centrifugal force that can
act on the vehicle and tries to push it to the outside of the
curve. Furthermore, higher speed may result in higher
centrifugal force.
Regarding the differences at other segments, CR at
merge segment increases rapidly at high-density stages and
gets much higher compared to basic and diverge segments.
The results of paired t-test at the three facility types in
Table 4 also reveal that CR at basic/diverge segments is
significantly lower than that at merge segment, while they
are not significantly different from each other between
basic and diverge segments. At merge segment, merging
maneuvers can result in slow-down and lane-changing
behaviors for mainline traffic. These interruptions may
increase the possibility of vehicle conflicts. Such possi-
bility can further increase with an increase in k.
Table 5 summarizes the CR regression models as
function of k as well as the goodness-of-fit of models at
four facility types. At sharp curve, the model is power
function while they are quadratic functions at other facility
types. All of the models and variables are significant at
95 % confidence level (not shown in Table 5). Regarding
quadratic models, CR at merge segment is most sensitive to
the increase in k, more than three times of CR increases as
that at basic and diverge segments by the increase in one
unit of k.
4.2 Congested flow
Figure 9 describes the differences of CR distribution to
k by facility type in congested flow. It appears that CR
follows increasing tendencies to k at four facility types. In
contrast to other segments, sharp curve still has the highest
CR in congested flow while no statistical regression model
is developed at this facility type due to the limited crash
samples. Since the differences of CR tendency at other
segments are not clear in Fig. 9, a paired t-test is conducted
as shown in Table 6. The results indicate that there is no
112
97
36 32 31
11
55
72
21 28
48
28
81 73
22 14
8 6
228
123
64
42 30 26
0
5
10
15
20
25
30
0
50
100
150
200
250
300
5 15 25 35 45 55
Cra
sh ra
te (c
rash
/106 V
KM
T)
Num
ber o
f cr
ashe
s
Traffic density (veh/km)
NOC at basic segmentNOC at merge segmentNOC at diverge segmentNOC at sharp curveCR at basic segmentCR at merge segmentCR at diverge segmentCR at sharp curve
Fig. 8 CR comparison in uncongested flow
Table 4 T-test of CR in uncongested flow
Paired t-Value df Sig.
Pair 1: basic and merge segments -2.781 5 0.019
Pair 2: basic and diverge segments -1.070 5 0.310
Pair 3: merge and diverge segments 2.320 5 0.043
Comparative study on crash-influencing factors 229
123J. Mod. Transport. (2013) 21(4):224–235
significant difference of CR at basic, merge, and diverge
segments. Such finding may imply that the effect of facility
type on crashes is reduced in congested flow. For this
reason, CR model is developed by combining the three
facility types in order to increase the number of crash
samples for reliability. As demonstrated in Table 7, an
exponential function is adopted and it fits well to the
combined CR tendency. The model and its variables are
also significant at 95 % confidence level, while the results
are not shown in Table 7.
5 Effects of influencing factors
The analyses above reveal that CR characteristics are
different by facility type, which may be related to the
different geometric designs and traffic characteristics.
However, CR analysis is insufficient to examine a variety
of factors by a single model. Instead, PCA is applied and
the affecting mechanisms of individual factors are further
investigated.
5.1 Introduction of variables
Table 8 explains individual variables combining with its
type and some summary statistics. In nature, traffic flow
diagram is two-dimensional, and k and v are used together
to describe traffic conditions. As for geometric features, h,
ICF, and ISD are picked out to reflect the vertical, hori-
zontal, and comprehensive geometric variations, respec-
tively. Dummy variables are referred to incorporate
ambient conditions into PCA. A dummy variable usually
takes 0 and 1. In this case, weather conditions (over 2
categories) are replaced by pavement conditions (only 2
categories), since two conditions are usually highly related
to each other.
At merge and diverge segments, ramp traffic is a sig-
nificant influencing factor on crashes [16]. This study
employs ramp flow ratio to illustrate the interaction
between ramp and mainline traffic. Merging ratio (MR) or
diverging ratio (DR) is defined as the proportion of on- or
off-ramp traffic out of the sum of ramp and mainline traffic,
respectively. Meanwhile, the length of acceleration lane LA
or the length of deceleration lane LD is adopted to reveal
the space available provided for merging or diverging
maneuvers, respectively.
Table 5 CR regression models in uncongested flow
Facility type Sample size Modela
B 319 crashes CR = 6.81 9 10-4k2-3.23 9 10-2k ? 0.541, R2 = 0.998, k(CRmin) = 24
M 25 l crashes CR = 2.55 9 10-3k2-9.29 9 10-2k ? 1.01, R2 = 0.983, k(CRmin) = 20
D 204 crashes CR = 5.68 9 10-4k2-3.34 9 10-2k ? 0.747 R2 = 0.849 k(CRmin) = 28
S 513 crashes CR = 132.2k-0.983, R2 = 0.992
a k (CRminimal): traffic density corresponding to the minimal CR
33
45
29
20
11
31
43
12
5
23
14
7 4
16
6 5
0
5
10
15
20
25
30
35
40
45
0
10
20
30
40
50
60
70
80
90
55 85 115 145 175
Cra
sh ra
te (
cras
h/10
6V
KM
T)
Num
ber o
f cr
ashe
s
Traffic density (veh/km)
NOC at basic segmentNOC at merge segmentNOC at diverge segmentNOC at sharp curveCR at basic segmentCR at merge segmentCR at diverge segmentCR at sharp curve
Fig. 9 CR comparison in congested flow
Table 6 T-test of CR in congested flow
Paired t-Value df Sig.
Pair 1: basic and merge segments -2.448 4 0.092
Pair 2: basic and diverge segments -0.153 4 0.888
Pair 3: merge and diverge segments 0.325 4 0.767
Table 7 CR regression model in congested flow
Facility type Sample size Model
B ? M ? D 513 crashes CR = 4.87 9 10-1e0.0155k R2 = 0.924
230 Y. Wu et al.
123 J. Mod. Transport. (2013) 21(4):224–235
5.2 PCA among various facility types
In essence, PCA rotates data by using a linear transfor-
mation. Consequently, only the monotonic loadings of
factors can be reflected by this approach. For this reason,
uncongested flow is further classified into low-and high-
density conditions at approximately 25 veh/km in view of
the value of k (CRmin) as shown in Fig. 10, since there are
different monotonicities of CR model in two conditions. As
a result, three traffic conditions are analyzed, i.e., low- and
high-density uncongested flow as well as congested flow.
5.2.1 Low-density uncongested flow
Table 9 demonstrates PCA results at basic segment in low-
density uncongested flow. In terms of the rules introduced
in Sect. 3.4, four components are selected and all of the
factors can explain at least 80 % of variance in the original
dataset in terms of the value of cumulative percent.
In low-density uncongested flow, crashes at basic seg-
ment are found to be significantly associated with geo-
metric variation (ICF and ISD), traffic density along with
ambient light, speed coupled with pavement, and vertical
variation h. Geometric variation is the 1st component, as
great variation may result in frequent speed reduction.
Accordingly, the difficulty for drivers to control vehicle
behaviors increases. At low traffic density k, driver’s
attention is not high, and some discretionary behaviors may
be operated. Such condition combining with the poor
ambient light is possible to increase crash risk. Meanwhile,
due to the reduced value of tire-pavement friction, high
speed v combining with wet pavement can reduce the
roadability. In such cases, k and v are two separate com-
ponents, which can further demonstrate that both variables
are not highly interrelated at low flow rate. In addition,
vertical variation h has a positive loading because of the
increased visibility restriction and the difficulty in main-
taining vehicle behaviors for drivers.
Principal components at other segments are analyzed as
shown in Table 10. The variables that are significantly
related to each component are selected based on their
loadings. For judging the relative significance of the same
component by facility type, the percent of variance
explained by each component is provided as well.
One difference at merge segment from basic segment is
that MR combining with the length of acceleration lane LA
becomes a principal component. Meanwhile, day type is
found to be significant. In terms of the percent of variance
accounted by components, the significance of geometric
variation gets lower in contrast to basic segment. Merging
traffic is an important influencing factor, since it can induce
interruption to mainline traffic. Such interruption may get
stronger as MR increases. Besides, higher LA can provide
more space for ramp and mainline traffic to adjust for
Table 8 Introduction of individual variables
Variables Statisticsa Description
Max. Min.
k 238 0 Traffic density (veh/km)
v 141.0 4.7 Average speed (km/h)
MR 0.78 0.00 Merging ratio
DR 0.60 0.00 Diverging ratio
ICF 1997 0.1 Index of centrifugal force (km3/h2)
h 14.8 0.1 Vertical variation (m)
ISD 1112.7 0.0 Index of space displacement (m2)
LA 250 100 Length of acceleration lane (m)
LD 432 100 Length of deceleration lane (m)
Pave f(1) = 27.7 % Equal to 1 if wet pavement, 0 otherwise
Light f(1) = 27.1 % Equal to 1 if nighttime, 0 otherwise
Day f(1) = 29.4 % Equal to 1 if holiday, 0 otherwise
f frequencya Max./Min.: maximal and minimal values, respectively
Congested flow
Uncongested flow
Low-density High-density
vc
0Traffic volume (q)
Spee
d (v
)
ke=2
5 ve
h/km
vC=60 km/h (basic/merge/diverge segments)vC=45 km/h (tight curve)
Fig. 10 Classification of traffic conditions
Table 9 PCA results at basic segments
Variables Component
1st 2nd 3rd 4th
ke -0.194 -0.852 -0.119 0.103
v 0.285 0.182 0.798 -0.086
ICF 0.953 0.005 0.053 -0.122
ISD 0.959 0.011 -0.044 0.090
h 0.119 0.149 0.228 0.973
Pave 0.294 0.214 0.783 -0.095
Day 0.139 0.093 0.190 -0.468
Light -0.164 0.838 -0.130 0.143
Initial Eigenvalue 2.12 1.54 1.37 1.12
Percent of variance 30.3 20.2 18.2 15.1
Cumulative Percent 30.3 50.5 68.7 83.8
The variables highly related to each component are in bold
Comparative study on crash-influencing factors 231
123J. Mod. Transport. (2013) 21(4):224–235
merging behaviors. Regarding the influence of day type on
crashes, it may be related to the different vehicle compo-
sitions and driver populations between holiday and week-
day, while such influence needs a further study to
investigate vehicle behaviors at merge segment. As for
geometric variation, on ramps in NEX are virtually allo-
cated far from poor alignment like small curve. Thus, it is
considered reliable that the significance of geometric var-
iation affecting crashes is lower at merger segment com-
pared to basic segment.
At diverge segment, the most significant difference from
basic and merge segments is that the DR and the vertical
variation h are related to the 1st component. Higher DR can
significantly interrupt mainline traffic since it is necessary
to pass through several lanes to move onto the deceleration
lane for driving vehicles. Furthermore, higher h can make
lane-changing maneuvers more difficult.
Generally, sharp curve has much worse design consis-
tency compared to other segments. Crashes at sharp curve
are found to be associated with poor vertical consistency
(ISD and h), high horizontal variation ICF along with speed
v, low traffic density k in nighttime, wet pavement, and
holiday. In NEX, sharp curve is often designed to connect
routes with different elevations. Thus, the vertical consis-
tency is fairly poor. Smaller radius along with high v may
cause notable centrifugal force. The affecting mechanisms
of other component are similar to these at basic, merge, and
diverge segments.
5.2.2 High-density uncongested flow
As traffic density increases, the inter-vehicle interaction
gets more intensive. The corresponding results of PCA in
high-density uncongested flow are summarized in
Table 11. All of the components are of statistical
significance.
In the case of high-density uncongested flow, it is dis-
tinct that traffic-related variables including k and v become
an independent component, as a reflection of the increased
interaction of vehicles. Furthermore, in terms of the value
of loading, high density not low density is adverse to
safety. The finding can further support the results of CR
models: CR is decreasing to k in low-density uncongested
flow, while it is increasing in high-density uncongested
flow.
With respect to the differences by facility type, at merge
segment, MR gets to be a factor related to the 1st com-
ponent due to the increased interruption of ramp traffic
with the increase of traffic density. During the variation in
traffic conditions, the significance of DR becomes lower
than geometric variation at diverge segment. However, in
high-density uncongested flow, LD is more important in
contrast to low-density uncongested flow. Once a driver
feels the difficulty for lane-changing maneuvers in
diverging area, they may move onto the nearest lane to off-
ramp in advance in the upstream of diverging area. As a
result, the impact of lane-changing maneuvers on mainline
traffic gets relatively low. In a sharp curve, crashes are still
found to be probable with a decrease in k, which is similar
to the tendency of CR model.
5.2.3 Congested flow
With the further increase of traffic density, congested flow
appears. In the same way, Table 12 summarizes the results
of PCA by facility type in congested flow.
Table 10 PCA results in low-density uncongested flow
Facility type (number of crash) Item Principal component
1st 2nd 3rd 4th 5th
F L F L F L F L F L
Basic segment (225) Component ICF 0.953 k -0.852 v 0.798 h 0.973
ISD 0.959 Light 0.838 Pave 0.783
Percent of variance 30.3 20.2 18.2 15.1
Merge segment (140) Component ICF 0.867 k -0.841 MR 0.808 v 0.721 Day 0.874
ISD 0.901 Light 0.869 LA 0.747 Pave 0.742
Percent of variance 20.1 18.8 15.4 13.9 11.9
Diverge segment (167) Component DR 0.807 ICF 0.948 k -0.807 v 0.892 LD 0.797
h 0.845 ISD 0.746 Light 0.860 Pave 0.848
Percent of variance 18.9 17.6 16.2 15.3 13.3
Sharp curve (319) Component ISD 0.854 ICF 0.948 k -0.820 Pave 0.929 Day 0.981
h 0.978 v 0.753 Light 0.794
Percent of variance 23.7 17.5 16.4 13.6 12.7
F factor, L loading
232 Y. Wu et al.
123 J. Mod. Transport. (2013) 21(4):224–235
Table 12 demonstrates that the effect of traffic flow on
crashes get more important in congested flow, compared to
that in uncongested flow. Except merge segment, the sig-
nificance of traffic flow affecting crashes is the highest.
Based on the percent of variance, the influence of geo-
metric design is further decreasing.
Regarding the differences by facility type, crashes at
merge segment are found to be positively associated with
smaller LA, not higher LA. For congested flow, smaller LA
may increase the difficulty of adequate speed adjustment
for merging and lane-changing maneuvers. Besides, based
on the loading of day type, weekday not holiday is a sig-
nificant factor. It is likely related to higher percentage of
heavy vehicles on weekday that may induce more frequent
shockwave in congested flow. At diverge segment, as
similar to merge segment, weekday is also a significant
factor. Meanwhile, smaller LD not higher LD is adverse to
safety. At sharp curve, poor ambient light can significantly
restrict visibility, while visibility is critical for driving in
small inter-vehicle spacing. Thus, ambient light becomes
another important factor in congested flow compared to
high-density uncongested flow.
From the analyses above, geometric features are found
out to be the most significant influencing factor in uncon-
gested flow. In this sense, the different CR characteristics
by facility type in uncongested flow may be significantly
associated with the variation in geometry. Poor design
consistency induced by small radius is the potential cause
for the highest CR in sharp curve. Ramp traffic can inter-
rupt mainline traffic, and longer acceleration lane may
provide longer interruption area. Both features can increase
crash risk at merge segment. A lot of diverging traffic may
move onto the lane nearest to deceleration lane in advance
in the upstream of diverging area, since urban expressway
carries a lot of commuters and many drivers are familiar
with road structure. Hence, even if DR and LD are found
out as significant influencing factors, CR at diverge seg-
ment is not significantly higher than that at basic segment.
As traffic density increases, the effects of traffic-related
variables increase and get more significant than geometry
in congested flow. In this condition, once a breakdown
initiates at bottlenecks, it can propagate to upstream section
that may consists of several facility types, where traffic
conditions are not significantly different. As a result, the
difference of CR characteristics at basic, merge, and
diverge segments is not significant. However, due to the
heavily restricted visibility induced by the special geo-
metric design, sharp curve still has higher CR than other
facility types.
6 Conclusions and future work
This paper identified the different CR characteristics by
facility type of Nagoya Urban Expressway. In uncongested
flow, CR at basic, merge, and diverge segments appears
convex downward to traffic density. In contrast, the value
at sharp curve follows a decreasing tendency. In congested
flow, CR at four facility types increases as traffic density
increases. In both un- and congested flows, sharp curve has
the worst safety performance in view of its highest CR
among the four facility types. As for other segments, merge
Table 11 PCA results in high-density uncongested flow
Facility type (number of crash) Item Principal component
1st 2nd 3rd 4th 5th
F L F L F L F L F L
Basic segment (94) Component ICF 0.983 k 0.858 Day 0.776 h 0.925
ISD 0.974 v -0.885 Light 0.781
Percent of variance 28.9 21.8 17.1 14.4
Merge segment (112) Component MR 0.818 k 0.936 LA 0.842 Day 0.810 Pave 0.963
ICF 0.923 v -0.899 Light 0.713
ISD 0.960
Percent of variance 28.3 18.0 13.5 12.3 10.4
Diverge segment (37) Component ICF 0.970 DR 0.904 k 0.793 h 0.733 Pave 0.913
ISD 0.977 LD 0.772 v -0.704
Percent of variance 21.8 19.8 16.8 13.7 13.5
Sharp curve (122) Component ICF 0.723 k -0.901 Pave 0.901 Day 0.869
ISD 0.962 v 0.872
h 0.908
Percent of variance 33.2 21.9 15.3 13.9
Comparative study on crash-influencing factors 233
123J. Mod. Transport. (2013) 21(4):224–235
segment has higher CR compared to the basic and diverge
segments in uncongested flow. Comparatively, CR at three
facility types is not significantly different in congested
flow.
The causes of the differences were further investigated
by focusing on traffic conditions and considering the
interaction of geometry, traffic flow, and ambient condi-
tions. Generally, geometric features are the most significant
factors in uncongested flow. With the increase of traffic
density, the effects of traffic-related variables increase and
become most significant in congested flow. For ramp
traffic, the significance of MR affecting crashes is on the
rise as mainline flow increases. In contrast, the significance
of DR gets decreasing. In addition, higher LA and LD are
adverse to safety for uncongested flow, while smaller LA
and LD are adverse for congested flow. Crashes at sharp
curve are highly associated with the after effects of its
special geometric design, such as large centrifugal force
and heavy restricted visibility.
The potential benefits of integrating these findings in
safer geometric design and traffic control are numerous.
The analysis can provide a basis for geometric audit for
safety regarding design consistency. Meanwhile, based on
the estimated CR models, road administrators can easily
image the safety performance with the variation of traffic
conditions at a given facility type. Furthermore, PCA
results may help prioritize countermeasures and further
estimate the safety performance of an adopted
countermeasure.
For more accurate analysis of crash characteristics, data
in smaller time window, e.g., 1 min even 30 s, are highly
recommended to improve the reliability of statistics, since
crash occurrence is significantly associated with the short-
term turbulence of traffic flow [1]. Furthermore, it is better
to examine the effect of inter-lane interaction on crashes if
the lane-based data is available. In this study, ramp traffic
is found out to play a significant role for safety at merge
and diverge segments. Thus, a microscopic analysis on
driver behavior is needed. In essence, PCA is a qualitative
analysis and the results are insufficient for applying spe-
cific countermeasures for a given case. Future studies are
expected to acquire the quantitative effects of various
influencing factors on crashes.
Acknowledgments The authors gratefully acknowledge the support
of Nagoya Expressway Public Corporation for the data provision and
other helps for this study.
Open Access This article is distributed under the terms of the
Creative Commons Attribution License which permits any use, dis-
tribution, and reproduction in any medium, provided the original
author(s) and the source are credited.
References
1. Abdel-Aty M, Pande A (2007) Crash data analysis: collective
versus individual crash level approach. J Saf Res 38(5):581–587
2. Wu Y, Nakamura H, Asano M (2012) A comparative study on
crash rate characteristics at different intercity expressway facility
Table 12 PCA results in congested flow
Facility type (number of crash) Item Principal component
1st 2nd 3rd 4th 5th
F L F L F L F L F L
Basic segment (138) Component k 0.950 ICF 0.842 h 0.699 Day 0.942
v -0.947 ISD 0.743 Pave 0.814
Percent of variance 26.0 20.3 18.9 15.4
Merge segment (91) Component MR 0.871 k 0.883 ICF 0.672 LA -0.871 Day -0.933
h 0.790 v -0.884 Light 0.745 Pave 0.716
ISD 0.825
Percent of variance 22.1 18.3 14.6 13.9 10.1
Diverge segment (48) Component k 0.868 ICF 0.849 DR 0.854 Pave 0.721 Light 0.951
v -0.801 ISD 0.879 LD -0.783 Day -0.716
h 0.859
Percent of variance 21.9 19.6 15.4 13.8 11.6
Sharp curve (27) Component k -0.772 ICF 0.909 Pave 0.799 Day 0.912
v 0.950 ISD 0.875 0.949
h 0.971 Light
Percent of variance 30.5 26.6 22.5 11.1
234 Y. Wu et al.
123 J. Mod. Transport. (2013) 21(4):224–235
types. In: Proceedings of the 46th IP meeting of Japan Society of
Civil Engineering (JSCE), Saitama University, 2–4 Nov 2012
3. Golob TF, Recker WW (2004) A method for relating type of
crash to traffic flow characteristics on urban freeways. Transp Res
Part A 38(1):53–80
4. Rengarasu TM, Hagiwara T, Hirasawa M (2009) Effects of road
geometry and cross section variables on traffic accidents using
homogeneous road segments. Transp Res Rec 2102:34–42
5. Shively TS, Kockelman K, Daminen K (2010) A bayesian semi-
parametric model to estimate relationships between crash counts
and roadway characteristics. Transp Res Part B 44(5):699–715
6. Bajwa S, Warita H, Kuwahara M (2010) Effects of road geom-
etry, weather and traffic flow on safety. In: Proceedings of the
15th international conference of Hongkong Society for trans-
portation study, The Hong Kong University of Science and
Technology, Hong Kong, 11–14 Dec 2010
7. Abdel-Aty M, Pemmanaboina R (2006) Calibrating a real time
traffic crash prediction model using archived weather and ITS
traffic data. IEEE Trans Intell Transp Syst 7(2):167–174
8. Caliendo C, Guida M, Pasis A (2007) A crash-prediction model
for multilane road. Accid Anal Prev 39(4):657–670
9. Christoforou Z, Cohen S, Karlaftis MG (2011) Identifying crash
types propensity using real-time traffic data on freeways. J Saf
Res 42:43–50
10. Hossain M, Muromachi YA (2012) A bayesian network based
framework for real-time crash prediction on basic freeway seg-
ments of urban expressways. Accid Anal Prev 45:373–381
11. Wu Y, Nakamura H, Asano M (2013) A comparative study on
crash characteristics between urban and intercity expressway
basic segments. In: Selected proceedings of the 13th word con-
ference on transport research, 15 pp
12. Hikosaka T, Nakamura H (2001) Statistical analysis on rela-
tionship between crash rate and traffic flow condition in basic
expressway sections. In: Proceedings of the 21st Japan Society of
Traffic Engineering Meeting, Tokyo, 29–30 Oc 2001. (in
Japanese)
13. Ng JC, Sayed T (2004) Effects of geometric design consistency
on road safety. Can J Civ Eng 31:218–227
14. Shawky M, Nakamura H (2007) Characteristics of breakdown
phenomena in Japan urban expressway merging sections. In:
Proceedings of TRB 86th annual meeting, Washington DC,
21–25 Jan 2007
15. Orme JG, Combs-Orme T (2009) Multiple regression with dis-
crete dependent variables. Oxford University Press, New York
16. Zhou HG, Chen HY, Zhao JG et al (2010) Operational and safety
performance of left-sides off-ramps at freeway diverge areas. In:
Proceedings of TRB 89th annual meeting, Washington DC,
10–14 Jan 2010
Comparative study on crash-influencing factors 235
123J. Mod. Transport. (2013) 21(4):224–235